1. Normalization of the Speech Modulation Spectra for Robust Speech Recognition Xiong Xiao, Eng Siong Chng, and Haizhou Li
Wen-Yi Chu
Department of Computer Science & Information Engineering
National Taiwan Normal University

2. Outline Introduction Noise Effects on Speech Modulation Spectra
Temporal Structure Normalization
Segment-based Implementation
Experiments
Conclusion

3. Introduction(1/2)
The modulation spectra of a speech signal are the power spectral density (PSD) functions of the feature trajectories generated from the signal, hence they describe the temporal structure of the features.
When speech is distorted by noise, the temporal characteristics of feature trajectories are also distorted and need to be enhanced.
We propose the temporal structure normalization (TSN) filter to reduce the noise effects by normalizing the modulation spectra to reference spectra.

4. Introduction(2/2)
In this paper, we conduct an inquiry into the theoretical and practical issues of the TSN filter that includes the following.
1)We investigate the effects of noises on the speech modulation spectra and show the general characteristics of noisy speech modulation spectra.
2) We evaluate the TSN filter on the Aurora-4 task and demonstrate its effectiveness for a large vocabulary task.
3) We propose a segment-based implementation of the TSN filter that reduces the processing delay significantly without affecting the performance.

5. Noise Effects on Speech Modulation Spectra Definition of Modulation Spectra
For a time domain speech signal , its short-time power spectral density is defined as
where and are the sample index, frame index, and the acoustic frequency index, respectively. The denotes the short-time Fourier transform (STFT), and denotes the magnitude.

6. Noise Effects on Speech Modulation Spectra Modulation Spectra of Clean Speech and Noises
As the modulation spectra are averaged over 200 utterances, the effects of the spoken content and speaker are reduced.
From Fig. (a), it is clear that the modulation spectra of noises are flatter than that of speech.

7. Noise Effects on Speech Modulation Spectra Modulation Spectra of Noisy Speech
The noisy modulation spectrum of any filterbank is represented as
The three noisy modulation spectra have higher power than the clean modulation spectrum in the high modulation frequency range.
The flatness of the noisy modulation spectra are inversely proportional to the SNR level.

8. Temporal Structure Normalization Overview of TSN Framework(1/2)
The TSN filter aims to reduce the temporal differences between the clean and noisy features, such as the smoothness of the feature trajectory.
The magnitude response of the filter is designed from the PSD functions of the feature trajectories.

9. Temporal Structure Normalization Overview of TSN Framework(2/2)
Specifically, the magnitude response is designed to modify the feature trajectory’s PSD function towards a reference PSD function. In this way, the temporal characteristics of the features are normalized.
In both the training and processing steps, the feature trajectories are preprocessed individually by either the mean and variance normalization (MVN) or HEQ before the PSD estimation.

10. Temporal Structure Normalization Filter Design(1/2)
The filter design process is divided
into two parts:
1) estimates the desired magnitude
response of the filter with optional
modification (steps 1–3 in Table I)
2) designs the time domain finite-
impulse response (FIR) filter that
implements the desired magnitude
response (steps 4–7 in Table I)
The same filter design process is performed individually for each feature trajectory.

11. Temporal Structure Normalization Filter Design(2/2)
To normalize to ,the filter’s magnitude response should be
During the filter design process, there are two considerations.
1) Normalizing the Trend of the PSD Function Only:
The simplest approach is to use a smooth and in the
filter design process. (the autoregressive (AR) model-based
Yule–Walker method instead of the Fourier transform)
2) Combination With Other Temporal Filters:
To take advantage of other temporal filters, we can simply
multiply their magnitude responses with that of the TSN filter
(step 3 of Table I)

12. Temporal Structure Normalization Training of Reference PSD Functions
In the TSN framework, the general temporal structure is obtained by estimating the reference PSD functions from averaging the PSD functions of feature trajectories over a larger number of utterances.
The averaging process reduces other factors affecting the PSD functions, such as spoken content and speaker’s characteristics.

13. Segment-based Implementation
When short processing delay is desired or environmental conditions change rapidly within one utterance, it is more desirable to process the features in a segment-by-segment fashion.
The segment-based MVN and HEQ are implemented using the same
segmenting scheme as the TSNseg.
Considerations:
Two important parameters of
the TSNseg are the segment
length and the segment shift.

14. Experimental Setup
The TSN filter is evaluated on both the small vocabulary Aurora-2 task and the large vocabulary Aurora-4 task.
Mel-frequency cepstral coefficients (MFCCs) are used as the features for speech recognition.
For the TSN filter, the Yule–Walker method is used to estimate the PSD functions of feature trajectories.
A filter length of 33 taps is used for the evaluation if not otherwise stated.

15. Evaluation on the Aurora-2 Task(1/2)
From the table, the performance of the TSN filter is similar to that of the ARMA filter and better than that of the RASTA filter.
From the table, it is found that the integrated case TSN+ARMA is consistently better than both the ARMA filter and the TSN filter alone.
With the combined magnitude response, the TSN filter not only normalizes the feature’s temporal structure, but also applies additional smoothing to the features.

16. Evaluation on the Aurora-2 Task(2/2)
To ensure that the results of these tests are statistically significant, we carry out hypothesis test as follows.
Evaluation on Different Features
The TSN filter alone may not produce the best performance. However, the integration of the TSN with the ARMA filter (TSN ARMA) always improves the performance.

17. Evaluation on the Aurora-4 Task
From table, the results of TSN are always significantly better than those of ARMA and RASTA.
While feature smoothing can reduce the feature mismatches, it also leads to removal of useful speech information.
The degree of smoothing is a tradeoff between these two factors and usually depends on the signal conditions such as the SNR level and the task complexity.

18. Segment-Based Implementation
the gain of TSNseg over TSN is
mainly due to its use of MVNseg
preprocessing.
The main advantage of the TSNseg
is the reduction of the processing
delay.
With the segment-based implementation, the TSN filter can adapt much faster to run-time changing environments, and is thus suitable for low-delay speech recognition deployments.

19. Effect of Filter Length
A long filter can model the desired magnitude response better, but also means higher computational cost and a longer transition period for filtering.
Due to these considerations, a short filter is preferred if it does not degrade the performance too much compared to a long filter.

20. Contribution of Modulation Frequency Bands
It is generally believed that the intelligibility of human speech is closely related to the 1–16 Hz modulation frequency range.
This suggests that the performance improvement of the TSN filter should mainly come from its gain in the 1–16 Hz range.

21. Conclusion The major contributions of the paper include the following:
It discusses the effects of additive noises on the speech modulation spectra.
It conducts a comprehensive evaluation of the TSN framework with respect to different parameter settings.
It proposes a novel segment-based implementation of the TSN filter.
The major advantage of the TSN filter is its ability to adapt to changing environments.
From statistical point of view, the TSN normalizes the correlation of features over different frames, while other normalization techniques such as HEQ normalize the probability distributions of the features.